Enhancing breast cancer detection in digital mammograms through hybrid HHO-CS, multi feature and triplet learning frameworks

Abstract

Breast cancer remains a significant global health concern, with a high incidence rate among women and a leading cause of cancer-related mortality. Early detection is paramount for improving patient health well-being, and mammography plays a central role in breast cancer screening. While Computer Aided Detection and/or Diagnosis (CAD) systems have bolstered cancer detection rates by assisting radiologists in mammogram interpretation, challenges per- sist in conventional detection methods, particularly in cases of false negatives, especially in dense breast tissue. Subjectivity in manual image interpretation also contributes to diag- nostic variability. In recent years, the fusion of medical science with artificial intelligence, specifically deep learning, has introduced a transformative approach to breast cancer detection. Deep learning, primarily leveraging Convolutional Neural Networks (CNNs), offers a promising solution to address the limitations of traditional detection methods. The integration of deep learning has led to the development of CADet systems that significantly enhance accuracy while reducing false positives and negatives. This thesis embarks on a comprehensive exploration of various phases aimed at ele- vating the precision and quality of mammographic image analysis techniques. The research work encompasses four distinct studies, each contributing to the advancement of breast cancer detection methodologies. In Study-1, an innovative approach is introduced for pectoral muscle segmentation, incorporating watershed transformation and regularization techniques with masking. Ex- perimental results on the MIAS database demonstrate a remarkable accuracy of 97.20%. Study-2 focuses on breast image segmentation techniques, enabling the detection of breast tumors through machine learning methods, particularly support vector machines. A novel hybrid model, HHO-CS MKSVM, is proposed, exhibiting an impressive accuracy of 94.08% on the DDSM dataset. 1In Study-3, we investigate into mammography-based breast cancer diagnosis, leverag- ing deep learning methods to enhance both segmentation and classification. A multifeature learning framework is developed, encompassing preprocessing, segmentation, feature extrac- tion, and classification phases. The model achieves accuracies of 95.43% and 95.82% on the INbreast and MIAS datasets, respectively, showcasing its effectiveness in improving breast cancer diagnosis. Study-4 represents an innovative fusion of Explainable AI, a Triplet Deep Learning model, and an Adaptive Deep Neural Network (A-DNN) to further enhance segmentation and classification. Impressively, it achieves an accuracy of 97.98% on the DDSM dataset, bridging the gap between advanced AI and medical diagnostics. Collectively, these studies highlight the potential of deep learning-based methodologies to revolutionize breast cancer detection and diagnosis. By addressing challenges in segmenta- tion, classification, and interpretability, these innovative algorithms offer promising avenues for improving the accuracy and efficiency of mammographic image analysis, ultimately con- tributing to the early detection and improved management of breast cancer.